Precision for your AI spend. See where every Claude Code dollar goes — down to the individual tool call — then feed it back to Claude to spend less. → caliper.run
- 100% local. Reads the transcripts Claude Code already writes to
~/.claude/projects. Nothing leaves your machine — no telemetry, no account, no cloud. - Open source, MIT. Every number is auditable: cache-aware estimates at real per-model rates, parity-tested against ccusage, formula shown in the UI.
- You're in control. Machine-wide totals down to the exact call that spent the money — and the tools to make the next session cheaper.
/plugin marketplace add Cost-Caliper/caliper
/plugin install caliper@caliper/caliper— launch the dashboard, pointed at your current session/optimize-spend— Claude reads your spend and, with consent, writes you a personalized cost-discipline skill/distill-fable— before Fable is removed from Claude Code, use it to analyze its own real past work on this machine and, with consent, write a personal skill teaching Opus to approximate its thinking style
Once installed, Caliper stays visible without you having to remember it's there:
- Session start — quietly makes sure the dashboard is running (reuses it if already up; starts one on a random loopback port otherwise — no browser is opened) and reports the URL, plus a recap of your last session's spend if there is one.
- Every turn — a throttled
💰 Session spend: $X.XXXX · agents: …reminder, shown only once spend has grown by a meaningful amount since it was last shown. - Session end — the dashboard is only shut down once no other active Claude Code session is still using it, so nothing leaks as an orphaned background process.
The reminder reuses the exact same cost reconstruction the dashboard itself shows — never a separate estimate — so the number in the reminder always matches the UI.
| Variable | Default | Effect |
|---|---|---|
CALIPER_DISABLE_HOOKS |
unset | Set to 1 to make all three hooks fully inert — no auto-launch, no reminders, no auto-cleanup. |
CALIPER_REMINDER_THRESHOLD_USD |
0.05 |
Minimum spend growth (in dollars) before the per-turn reminder shows again. |
Upgrading from workflow-lens (pre-0.24)? Renamed — once per machine:
/plugin marketplace remove workflow-lens
/plugin marketplace add Cost-Caliper/caliper
/plugin install caliper@caliper| Sessions by folder & day | Session drill-down |
|---|---|
 |
 |
- Machine-wide analytics — all-time spend, daily charts stacked by model, every folder ranked (with filter + sorts), cache economics.
- Session forensics — waterfall of the main chat + every subagent on the real time axis, subagents ranked by cost, and per-run workflow timelines split into inference vs tool time.
- "Nerfed by Fable" tracking — counts every time Fable 5's safety classifier declined a request or Claude Code re-served it on the fallback model, split main chat vs subagents, with red markers on the daily chart and one-click prompts to analyze the reasons or disable auto-fallback. (Per-step prompt drill-down:
/legacy/for now.) - The loop — scope-aware ⧉ Optimize spend buttons (machine / folder / session) copy your real numbers and live API pointers into a prompt; Claude analyzes the spend or writes you a cost skill.
- Guided tour & dark mode — ✦ Tour explains every panel on your own data; ☾ toggles a full dark theme.
- Self-updating — checks this repo and offers one-click updates. (Previous UI preserved at
/legacy/for one release.)
Screenshots are bundled demo data (node scripts/demo-data.mjs). Your dashboard shows your own transcripts, locally.
cost = input×in + cache_write_5m×in×1.25 + cache_write_1h×in×2.0 + cache_read×in×0.10 + output×out
Per-model rates (fable-5 $10/$50 · opus $5/$25 · sonnet $3/$15 · haiku $1/$5 per Mtok), streamed duplicates deduplicated by request. Estimates, not invoices — the UI says so wherever a dollar appears.
| Path | What it is |
|---|---|
packages/control-tower/ |
The dashboard: local web app + JSON API reconstructing sessions, workflows, and subagents from transcripts. |
packages/workflow-lens/ |
Workflow-file toolkit: parse, lint, instrument, replay, estimate (keyless CLI + library). |
commands/, skills/ |
/caliper (+ /control-tower alias), the caliper skill, /optimize-spend, /distill-fable. |
hooks/, scripts/hooks/ |
Automatic SessionStart/Stop/SessionEnd hooks (see above). |
scripts/ |
Session-aware launcher, demo-data generator. |
Run from source:
node scripts/demo-data.mjs
WFLENS_PROJECTS_ROOT=/tmp/caliper-demo/projects PORT=8912 node packages/control-tower/server.mjsTests: npm test in each package — keyless fixtures, zero API calls.
First Caliper tool. Next: smart routing to open models, then models fine-tuned on your own usage — caliper.run.
MIT